Unmixing 4-D Ptychographic Image: Data Approach
TimeTuesday, July 246:30pm - 8:30pm
DescriptionThis study focuses on unmixing 4-D ptychographic images which incorporate different linear combinations of basic modes. Least square method for this problem gives poor results. In this study, a machine learning method is proposed to achieve better accuracy. The training data, instead of being collected from experiments, are generated synthetically. Performances of different data generation methods are compared and the decreases of the cost function are shown. The neural network is tested with all data we have and the result is satisfactory. The algorithm is implemented on C, with LAPACK for CPU code as well as MAGMA for GPU code. Generally the CPU code works well with smaller number of training examples, while GPU code is faster when the matrices are larger. The nature of this algorithm inspires me to work out an algorithm to compute the matrix inverse based on neural network. The convergence of the mentioned algorithm is well studied in this paper.